Detecting Operational Adversarial Examples for Reliable Deep Learning